I should keep this in mind whenever I want a student to self-assess if he completed the task at hand, or not.

I like the way you highlight the main difference.

I haven't done in-depth production-level coding, but another way to assess their difference is that a lot of coding involves applying the same principles in different settings. My colleague dismissively calls it monkey-coding, but ever since I tried my hand at it, I don't quite agree, as it is not easy and requires a lot of complex thinking to do it properly. I remember @k00b calling out the AI-slop open-source contributions they were getting, qualifying it as spam.

But with this logic in mind, I guess vibe coding can still be qualified as correct if it carries out the task, as dirty as the code may be. However, vibe physics, at least the way that VC-guy was describing it, would require going into the unknown... something LLMs, in their current implementation, by definition can't really do.

To be fair, for many of the papers that get published, they don't really go into the unknown, and it's often the application of known principles in new materials, for instance. In that case, I guess vibe coding could become a thing, at least in the right hands?